From c38563749f790f7143a7aeca7c6d4ad0a7e74ed1 Mon Sep 17 00:00:00 2001
From: JWLHS <847135749@qq.com>
Date: Tue, 14 Jul 2026 04:13:00 +0800
Subject: [PATCH] fix: preserve linear bias during INT4 reconstruction

---
 comfy/quantization/torchao/quantize.py | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/comfy/quantization/torchao/quantize.py b/comfy/quantization/torchao/quantize.py
index 673672808..868a945c1 100644
--- a/comfy/quantization/torchao/quantize.py
+++ b/comfy/quantization/torchao/quantize.py
@@ -37,7 +37,6 @@ def quantize_model(
     return model
 
 
-# torchao plain_int32 safetensors suffixes
 _QUANT_SUFFIXES = (
     ".weight_scale", ".weight_zp",
     ".weight_b0", ".weight_b1",
@@ -89,6 +88,8 @@ def reconstruct_int4_state_dict(
         for suffix in _QUANT_SUFFIXES:
             sd.pop(f"{base}{suffix}", None)
 
+        bias = sd.pop(f"{base}.bias", None)
+
         in_f = qdata.shape[1] * 8
         out_f = qdata.shape[0]
 
@@ -99,12 +100,15 @@ def reconstruct_int4_state_dict(
             scale=scale.to(torch.float16),
             zp=zp.to(torch.int8),
             block_size=(b0.item(), b1.item()),
+            bias=bias,
         )
 
         if device.type != "cpu":
             layer._qdata = layer._qdata.to(device)
             layer._scale = layer._scale.to(device)
             layer._zp = layer._zp.to(device)
+            if bias is not None:
+                layer.bias.data = layer.bias.data.to(device)
 
         replaced[base] = layer